2025 Volume 25 Issue 7 Pages 7_1-7_19
Apart from seismic waveforms obtained by seismographs, one data source that records seismic motion and damage with a time axis is footage from shop cameras. The National Research Institute for Earth Science and Disaster Resilience (NIED) is continuing to archive video footage of customer-attracting facilities that have suffered significant earthquakes through a collaboration agreement with Aeon Co. By using this archive, we clarify the relationship between the images and the damage. It is hypothesized that this will facilitate more efficient and labor-saving responses and countermeasures based on the available evidence. In this study, a neural network model for discriminating the presence or absence of anomalies by machine learning was constructed and evaluated, using the videos taken by shop cameras during the shaking of an earthquake as the data source. The objective of this paper was to examine the fundamental principles of machine learning and to gain an understanding of a technique that can be employed to enhance accuracy. Consequently, the proposed method, which employed the concept of one-hot encoding and pre-processing, yielded an 84 % classification accuracy for earthquake images that were previously unknown to the model.
